A Comparative Approach to Understand the Performance of CMIP6 Models for Maximum Temperature near Tropic of Cancer Using Multiple Machine Learning Ensembles

Choosing appropriate climate models is key for enhancing our understanding of Global Climate Models (GCMs) and mitigating their inherent weaknesses in climatic applications. This study fills a knowledge gap by exploring model selection and evaluation near the Tropic of Cancer, focusing on using multiple ensemble approaches to reduce uncertainties in Coupled Model Intercomparison Project 6 (CMIP6) GCMs. The goal is to guide climate researchers in making informed modeling decisions by comprehensively assessing various machine-learning (ML) models for predicting maximum temperature (Tmax) using historical India Meteorological Department data and 22 CMIP6 models. Tmax is crucial as it influences heat-related stress on agriculture, human health, and water resources, making its accurate prediction vital for climate adaptation strategies. This study evaluates the performance of 14 multi-model ensembles (MMEs) based on ML techniques, alongside multi-ML ensemble (MMLE) approaches through rigorous cross-validation and hyperparameter tuning. Key findings show that K-nearest neighbors (KNN) and stacking models outperform others, showcasing superior predictive accuracy within the MMLE framework. The ensemble approaches outperform individual models, with MMLE achieving R² of 0.934, significantly higher than individual ML models except for KNN (R² of 0.933). Detailed visualizations and error analyses validate results, highlighting the strengths and weaknesses of models. This research highlights the significance of model selection and tuning in temperature data analysis, offering a robust methodology for model evaluation. The findings improve climate predictions, guide adaptive strategies for climate change, and support efforts in climate resilience and policymaking.